I love chemistry. I like the idea that a hundred or so elements is all you need to make up every physical object in the universe. The elements combine in defined ways to form molecules and life itself. Thus, you can decompose everything in the universe to these elements. Not only that, you can further decompose the elements themselves. There is a finite set of properties that can combine to form these very same elements.
The premise of this chapter is that there is a similar reductive structure to the intentional world, which the chemists have specified for the physical world. We are trying to define those elemental goals and their properties.
We expand the goal taxonomy of Schank and Abelson to include fifteen types of goals. We next present fifteen features or dimensions useful for characterizing goals. We discuss the use of importance in comparing goals and deciding between goals. We suggest that importance provides a principled alternative to the use of scales in cognitive modeling. We review the phenomenon of goal conflict, and suggest that goal conflict is inevitable and naturally gives rise to assigning importance to goals. We discuss the need for knowledge of stereotypical or default goals for various types of agents. We suggest that functional objects, such as automobiles and computers, have stereotypical goals associated with their use.
It is appropriate for us to discuss what we mean by a goal, since neither common usage of the word nor AI terminology has arrived at a uniform meaning. Our view of goals is quite broad, and should encompass most of the accepted interpretations of the word. We suggest the following general definition of a goal.
Definition 3.1 A goal is a state of the world which an agent explicitly desires to achieve, preserve, avoid, or destroy.
This definition is actually a starting point — a theme on which we will offer variations. We present the following types of goals and meta-goals as indicative of the scope of our intended use. This classification scheme is not meant to be orthogonal. A given goal may fall into more than one category. Furthermore, we make no claims that this is an exhaustive list. Rather, it is an attempt to illustrate the wide range of goal related phenomena.
We begin with the goal taxonomy of Schank and Abelson (1977). The Schank and Abelson goal system was intended primarily for the purpose of story understanding. This taxonomy comprises the first seven categories listed below.
Example 3.1 John was hungry.
Example 3.2 John needed to sleep.
Example 3.3 John was parched.
These commonly are physiological needs, such as the need for food and sleep. These goals are in effect not intentional but biological in origin. You cannot simply decide that you do not need to sleep. You do not have control over these goals. You may decide not to sleep, but you cannot decide that you are not sleepy. Appetites and addictions are types of S-goals. Under this definition, the desire for a member of Congress to seek reelection is not an S-goal. However, political scientists may wish to subject this assertion to empirical verification.
Example 3.4 John read a novel.
Example 3.5 John hiked in the woods.
Example 3.6 John played backgammon with Mary.
The agent is pursuing a goal for intrinsic enjoyment or relaxation. These are goals posited to explain why people go to concerts and watch TV. Pursuing an E-goal suggests that more “serious” goals are currently satisfied. For example, you do not spend money on a vacation to Europe if you need that money for a heart operation. Someone obsessed with worry would not pursue E-goals.
Schank and Abelson note that various S-goals could also be pursued as E-goals. That is, S-HUNGER could coexist with E-EATING. The satisfaction goal S-DRUG-ADDICTION could be associated with E-RECREATIONAL-DRUGS.
Example 3.7 John graduated from college.
Example 3.8 John cashed his paycheck.
Example 3.9 John bought a six-pack.
This is a general category for achieving some new state of affairs. Thus, buying groceries or being elected mayor are both A-goals. In our present framework, an A-goal can also be viewed as the acquisition of a resource.
A-goals are desires to obtain a certain — often persistent — state of the world. The typical goals in AI planning domains are A-goals. For example, a blocks world goal of building a tower would be an example of an A-goal. The goal of a robot to carry a box from one office to another would be an A-goal.
Example 3.10 John brushed his teeth every day.
Example 3.11 John had a physical exam.
Example 3.12 John put a second coat of paint on his house.
These are goals to maintain the status quo, such as your health, home, and positive relationships with friends and family. Changing the oil in your car or going to the dentist involve P-goals. A-goals, once achieved, are often followed by P-goals. Once you buy a house, you have to keep it up, e.g, P-HOME-INSURANCE, and P-REPAIRS.
AI planning programs may introduce P-goals. The blocks world planner may have a P-goal not to knock down the tower while it is under construction or after it is completed. The mobile robot may be instructed not to drop the box while carrying it from one room to another.
Example 3.13 John called the police to report the mugging.
Example 3.14 John grabbed the fire extinguisher.
Example 3.15 John pushed Mary away from the oncoming truck.
These usually involve a threatened P-goal. These are goals of sudden onset and high priority. When your house is on fire or you lose your job, you face a C-goal.
C-goals have a negative connotation. That is, if the agent does not take immediate action, something bad will happen.
Example 3.16 John dialed the phone.
Example 3.17 John opened the newspaper.
Example 3.18 John put the key in the ignition.
Many goals are pursued merely as instrumental to other goals. The higher level goals are the purpose or result or motive or aim. The sub-goals are important insofar as their result is important. This view of goals focuses on the ends of a course of action. The importance of an I-goal increases both if the higher level goal is important and if there are few or no alternatives to the I-goal. For example, John’s I-goal of starting the car becomes important if John has to drive his wife to the emergency room, and if there are no other means of transportation. John’s I-goal is less important if it is merely in service of John driving around the corner.
As described in Schank and Abelson, I-goals generally involve steps in a script or stereotypical plan. Thus, dialing a phone is part of the normal plan for making a phone call, and opening the newspaper is required for reading the newspaper. Clearly, there could be other reasons for these behaviors. John could be dialing the phone as physical therapy, and opening the newspaper to line a bird cage.
Example 3.19 John wanted to borrow Mary’s car.
Example 3.20 John wanted to know what movies were playing.
Example 3.21 John had to go to work.
Delta goals are similar to I-goals, but are posited to account for more general planning situations. That is, there is no particular script associated with D-goals. The three I-goal examples in the preceding section could serve as instantiations of plans for achieving the respective D-goal examples given here. Thus, John could call Mary on the phone to ask her to borrow her car (D-AGENCY). John could look in the newspaper to find out what movies were playing (D-KNOW). John could drive to work (D-PROX). There could be many other ways in which one could achieve these D-goals.
This completes the goal taxonomy of Schank and Abelson. It should be clear that these categories are not orthogonal. It is not difficult to classify a given goal in several ways at once. We shall now suggest several additional goal categories in an effort to illuminate other features of goals. We shall then propose explicit dimensions of goals to account for the various observed goal types.
Example 3.22 John bought a lottery ticket.
Example 3.23 John hoped the Mets would win the pennant.
Example 3.24 John prayed for peace.
People often express desires which cannot be directly fulfilled. The people have no direct plans or resources for achieving these goals. They are merely stating that they would be happy if circumstances conspired to satisfy this goal.
Example 3.25 John worried about getting sick.
Example 3.26 John hated the Pirates.
Example 3.27 John was concerned about inflation.
These are the opposite of wishes and hopes. These focus on possible negative outcomes that should be avoided. As with wishes, the agent may have no plans or resources for preventing the undesired outcome.
One might reasonably suggest that hopes and worries are the same phenomenon, i.e., hoping for a good outcome is the same as worrying that the good outcome will not occur. In chapter 4, we discuss the affects that are associated with goals and note that there are different emotional states associated with hopes and worries. Furthermore, in chapter 8 we discuss Prospect Theory which suggests that people make distinctions between seemingly identical outcomes determined by the emphasis on the positive or negative consequences.
Example 3.28 John had a flat tire.
Example 3.29 John’s shoelace broke.
Example 3.30 John could not remember the name of his boss’s wife.
When we are executing a plan and encounter a blocked step or missing resource, that’s a problem. It need not be a crisis. An accident is actually a non-goal; it is the unintended side-effect of an action. The importance of a problem or accident depends on the importance of the primary goal involved, and the availability of alternatives.
Example 3.31 John read People Magazine while waiting in the check-out line.
Example 3.32 John asked the babysitter to mail a letter for him on her way home.
Example 3.33 John asked the policeman for directions after he got a speeding ticket.
Certain plans can be expanded in simple ways to accomplish additional goals. These additional goals were not part of the original plan, but an agent can recognize opportunities for using an unexpected resource. Opportunities can be viewed as the opposite of problems. However, unlike problems, opportunities may be harder to identify [Birnbaum, 1986]. Problems normally present themselves as failures of expectation.
To the extent that we perceive additional resources as failures of expectations, we can recognize opportunities. Consider two more examples.
Example 3.34 John’s afternoon meeting was cancelled.
Example 3.35 John won the lottery.
In the first case, John had expected to spend the afternoon in a meeting, but now has the afternoon free. He should recognize this as an opportunity: A resource that had previously been committed is now available.
Winning the lottery is a similar case. Even though John had wanted to win, he did not expect to win the lottery. When John wins, he must reassess many of his plans and priorities. He may no longer have to work, and he can now purchase things that he could not previously have afforded.
Example 3.36 John is a Democrat.
Example 3.37 John is always polite.
Example 3.38 John is tight with a dollar.
Some goals can never be satisfied once and for all, but they can influence local behavior. In a sense, values and principles are meta-goals — they constrain what goals can be pursued. For example, if a person places great importance in honesty, he will not adopt plans that require him to behave in a dishonest fashion. Notice that the goal HONESTY is quite different from the goal HUNGER. HUNGER says “try to find a plan which will result in satisfying me.” HONESTY says “do not engage in any plans that violate me.” This has the flavor of a P-goal, but without any tangible object.
Wilensky (1984) has posited meta-goals such as “achieve as many goals as possible” and “do not pursue impossible goals.” His meta-goals tend to focus on the planning process itself. Wilensky’s meta-goals are independent of the content or semantics of the goals themselves. The values that we wish to posit as a type of goal focus on cultural knowledge, not particular planning knowledge.
We take issue with the formulation of Wilensky’s meta-goals, insofar as they fail to take into account the relative importance of disparate goals. For example, a planner should not adopt a plan to achieve as many goals as possible if that plan fails to achieve a very high level goal for which an effective plan exists. Similarly, it may be appropriate for an agent to adopt a plan that entails the pursuit of an impossible goal if the plan also achieves other goals. For example, someone could run for political office even though legally prohibited from serving (due to age or other restriction) simply to gain publicity or recognition.
Example 3.39 John filed his income tax return early.
Example 3.40 John ran a red light while taking his wife to the hospital.
Example 3.41 John was fined for being late to work three days in a row.
Laws are explicit principles of a society or organization. They act in the same way as values or principles in governing behavior. In addition, laws often have explicit means of enforcement and punishment associated with noncompliance.
Example 3.42 John has a penchant for red ties.
Example 3.43 John likes modern art.
Example 3.44 John prefers red wine to white.
Certain choices that people make have less consequence than other choices. Deciding whether to have the Dover sole or the steak tartare is probably less important than deciding whether or not to fast. However, matters of taste involve the same type of decision-making processes that apply to matters of moment. Preferences and tastes reflect a set of priorities. Some people are said to have “expensive tastes.” According to the principle of importance, this implies that such people are willing to expend greater resources to satisfy their sense of style.
Example 3.45 John worked late Friday night.
Example 3.46 John stayed home and took care of the baby.
Example 3.47 John served on the jury.
Certain goals arise from roles that a person adopts. For example, a parent has responsibilities that give rise to goals that require the parent to expend resources.
We conclude this taxonomy of goals with a disclaimer. We do not mean to suggest that this set of goal types is either necessary or sufficient. We merely wish to arrive at a working vocabulary for examining goal-based phenomena. We have chosen to expand the taxonomy of Schank and Abelson. Others have developed theories of goals and motivation. For example, Maslow (1970) presents a “holistic-dynamic” theory of human motivation that encompasses physiological needs, safety needs, belongingness needs, esteem needs, and the need for self-actualization — all considered “basic needs.” The system of classification is not an end in itself. It should provide us with a better way to identify and examine the underlying phenomena.
A planner must make a choice between alternatives if plans have a resource in conflict. For example, if a child wants two candy bars, A and B, but only has enough money for one, she must make a choice. The goals themselves are not in conflict. It is the limited resource of money that gives rise to the conflict. Thus, if the child had more money, she could simply buy both candy bars. Limited resources create conflicts that drive decisions.
In choosing among alternative plans, a decision maker will select the plan that best satisfies his goals. This statement seems straightforward, but it raises some questions. First, goals may be incomparable, so there may be no one best plan. Second, what can be properly considered a goal?
We observe that the answer to the first question in many cases may be undecidable. That is, within a given set of alternatives, different choices may be considered best depending on the method of ranking.
For example, Wilensky (1983) has suggested that the meta-theme achieve as many goals as possible (AAMGAP) could be used to resolve goal conflicts. That is, given two or more alternatives, choose the one which achieves the largest number of goals. Another meta-theme for resolving goal conflict is maximize the value of the goals achieved (MTVOTGA).
Consider the case of a Congressman who must choose between funding a space station or a range of social programs for veterans, the elderly, the handicapped, the homeless, and urban renewal (cf. HR-4800, June 22, 1988). We may assume that the social programs would satisfy numerous goals for many people, while the space station may have a greater intrinsic value. Using AAMGAP, the Congressman would vote for the social programs. Under MTVOTGA, he would opt for the space station. (Congress chose the space station by a vote of 256 to 166.)
One might argue that the space station could in fact be selected under AAMGAP as well, since a successful space project might be instrumental to the satisfaction of many other technological and even social goals. We become faced with two accounting questions: What counts as a goal and how much does a goal count?
Given our task of making and justifying decisions, we accept as a goal almost anything that could enter into the justification or explanation of a decision. In traditional AI models of explanation [Winograd, 1972, Swartout, 1983, Ellman, 1989], the answer to the question “Why” usually involves the traversal of a causal chain of goals with tight instrumental coupling. We observe that in political discourse, as well as in informal reasoning, explanations are not so often tightly reasoned chains of inference, but rather loosely connected associations.
Instances of our additions to the Schank and Abelson goal taxonomy can be found in the political domain of roll call voting.
Wishes and hopes. Achieving world peace.
Worries and concerns. Avoiding a nuclear holocaust.
Problems and accidents. Dealing with the AIDS epidemic.
Opportunities. Exploiting the end of the cold war to reduce defense spending.
Values, ideals, and principles. Bringing honesty and integrity to government.
Laws, rules, and taboos. Abolishing apartheid.
Tastes and preferences. Supporting gay rights.
Duties and responsibilities. Providing funding for the United Nations.
In answer to our first accounting question, what counts as a goal?, we consider the categories presented earlier as indicative of the broad range of heterogeneous goals. The second question, how much does a goal count?, requires that we adopt a scale for measuring and comparing different goals. We postulate that goals are ranked by importance according to the following principle.
Principle of Importance. The importance of a goal is proportional to the resources that the agent is willing to expend in pursuit of that goal.
That is, the relative importance of goals is determined when goals compete for the same resource. The more important goal is allocated the resource, all other things being equal. Importance is analogous to Wilensky’s value [Wilensky, 1983]. Importance may vary among agents, and even within one agent over time.
Importance is one facet of goals. There are more. Given the types of goals discussed above, we suggest that the following features are appropriate for describing and analyzing goals. We note that Owens (1988, 1990) among others has focussed on developing descriptive vocabularies for goals and plans with particular emphasis on memory organization and indexing. Figure 3.1 categorizes the above types of goals along these dimensions. We advise the reader to note that the entries in Figure 3.1 have a wide variance. Given that the goal types are not orthogonal, a given goal might fall into several categories, with different default dimensions. There are usually exceptions to generalizations — including this one.
Importance. We can prioritize goals to choose among alternatives. Importance provides a means of comparing and evaluating choices.
Figure 3.1 addresses the question “What level of importance is normally associated with this type of goal: low, medium, or high?”
Polarity. Goals can be positive or negative. The achievement of a positive goal is termed good, and a negative goal bad. If two positive goals have different importance, one is better than the other. In most cases, the negation of a goal has the opposite value. Generally speaking, a goal “to do something” is positive, whereas a goal “not to do something” is negative.
Figure 3.1 addresses the question “Is the polarity normally associated with this type of goal positive or negative?”
Source or Origin. Some goals, such as hunger, are biological in origin. Other goals, such as choosing Dover sole, are volitional. Physiological states lead to biological goals. Experience and beliefs lead to intentional goals.
Figure 3.1 addresses the question “What is the normal source associated with this type of goal: biological state, personal belief, circumstances (“circ”), or social custom (“cust”)?”
Intrinsic vs. Extrinsic. Some goals involve things done for their own satisfaction, such as watching a movie or reading a novel. These are goals of intrinsic satisfaction. We can term them desires. Other goals — most goals — are part of plans that are instrumental to the achievement of other goals, such as going to work or shopping for food. There may actually be a spectrum of motivation from completely intrinsic at one extreme to totally extrinsic at the other. Most goals will fall somewhere in between.
Figure 3.1 addresses the question “Is this type of goal normally of intrinsic or extrinsic satisfaction to the agent?”
Satisfaction Conditions. The states required to satisfy a goal and the common plans that may be used to achieve those states. Many goals will index the plans that are likely to be used to achieve those goals.
Figure 3.1 addresses the question “Are there typically many or few ways to satisfy this type of goal?”
Status. The state of the goal, e.g., active, satisfied, pending. When your hunger goal is active, you are hungry. After eating, your hunger goal is satisfied, but also pending, since you will later become hungry again.
Since status refers to the stage of action associated with a goal, there is no typical status for a given type of goal. Therefore, Figure 3.1 does not contain entries for goal status.
Likelihood. The chances of successful achievement of the goal. If John wants to go to the beach, but knows that it is likely to rain, he should try to pursue some other goal.
Figure 3.1 addresses the question “Is this type of goal likely to succeed (+) or fail (–)?”
Consequences and Outcomes. These come in two flavors. First, there are the results following the successful achievement of the goal. That is, the achievement of this goal enables other goals. Going to the grocery store is instrumental to purchasing food which is instrumental to cooking. Second, there are the results of failing to achieve the goal. For the goal of paying income tax, the failure consequence includes fines, interest, and a possible jail term. The success consequence is the peace of mind of avoiding the failure consequence. Each type of consequence involves cost and benefit.
Figure 3.1 addresses the question “For this type of goal, which outcomes matter most: success (+) or failure (–)?”
Deadline. A date or time by which a goal must be satisfied to assure the success consequence and avoid the failure consequence. You need to get to the grocery store before it closes. You need to file your income tax return before April 15th.
Figure 3.1 addresses the question “Does this type of goal typically have a deadline?”
Frequency and Persistence. How often does this goal arise and how long does it stick around? Physiological goals such as hunger and sleep have clear daily time courses. Other goals, such as raising children or pursuing a career, may persist for years and decades. There are normative frequencies for common goals. We would think it odd if a person bathed 30 times a day or just once a year.
For frequency, Figure 3.1 addresses the question “How many times per year might this type of goal occur: 1, 10, 100, 1000, or 10000?” For persistence, Figure 3.1 addresses the question “How long does a given instance of this type of goal typically persist before it is satisfied or abandoned: a minute (“min”), an hour, a day, a month, or a year?”
Age. The age of a goal indicates how long it has been present. This is different from persistence, which indicates the expected lifetime of a goal. For example, C-goals usually have sudden onset, which is to say, low age. However, C-goals may vary quite a bit in persistence; they may be acute or chronic. If a goal has persisted, it clearly is an old goal. We may term such a positive, chronic goal a longing.
Age, like status, reflects an aspect of processing particular goals. Therefore, Figure 3.1 does not contain any entries for the typical age of various types of goals.
Contexts and Triggers. A context is the set of conditions under which a goal normally arises. These trigger conditions may include events, dates, or states. For example, S-SLEEP normally occurs late in the evening or after ingesting sedatives. S-HUNGER normally occurs several hours after the preceding meal. Some goals may be triggered by the presence of their enabling resources, or the failure of higher level goals. For example, if John’s Saturday tennis game gets rained out, he might decide to watch baseball on TV. The resource of an uncommitted weekend afternoon can lead John to develop a goal. This is an example of opportunistic planning.
Figure 3.1 addresses the question “Does this type of goal occur in almost all contexts, or merely in some?”
Adopted Goals. Some goals are vicarious or by proxy. A Mets fan has adopted the goal of wanting the Mets to win. A parent feeds an infant. The parent has adopted the child’s hunger goal.
Figure 3.1 addresses the question “Is this type of goal likely to be an adopted goal?”
Serious vs. Frivolous. Serious implies that failure is to be avoided. The failure consequences force this attitude. A serious, purposeful goal is one that is more important than a frivolous goal. E-goals are frivolous. It does not matter if you fail to finish reading a novel or watching a TV show. C-goals are serious. Failing to treat a heart attack or save a drowning person can be fatal.
Figure 3.1 addresses the question “Is this type of goal likely to be a serious goal?”
Substitution. Alternative goals that may be pursued in lieu of the current goal. This is often a good measure of goal importance. That is, higher level goals are less likely to have alternatives. John may be willing to eat pizza instead of tuna salad, but it is less likely that he will want to marry Jane if he is unable to marry Mary, his heart’s desire.
Figure 3.1 addresses the question “Is this type of goal likely to substituted by another goal?”
Several of these dimensions impinge on importance. We now review the dimensions that have a bearing on determining a goal’s importance.
Polarity. We do not suggest that positive goals are inherently more important than negative goals. Nor do we believe that the reverse is true. However, we do consider the possibility that for some people, the polarity of a goal may influence its importance. That is, some people prefer to pursue positive goals (achieving good outcomes), and other people chose to focus on negative goals (avoiding bad outcomes). Prospect Theory, discussed in chapter 8, suggests that many people ascribe greater importance to avoiding a loss than to achieving a gain.
Example 3.48 John used the insurance premium money to buy a present for Mary.
Source or Origin. In general, goals of biological origin will take precedence over volitional goals. Pain and pleasure are formidable motivators.
Example 3.49 John was too tired to go to the movies with Mary.
Intrinsic vs. Extrinsic. Goals of intrinsic satisfaction are more likely to be pursued than goals of extrinsic motivation. An agent will expend more resources doing what he wants to do rather than what he is told to do.
Example 3.50 John watched the game on TV instead of raking the leaves.
Satisfaction Conditions. The number of plans available for achieving a goal and the difficulty or cost associated with those plans may affect the importance of goal. That is, a goal that is extremely expensive to achieve may result in a diminished importance for that goal.
Example 3.51 John had only five minutes to get to work, so he skipped breakfast and planned to catch an early lunch instead.
Status. An active goal should have a higher priority than a pending goal, all other things being equal. It is usually more important to allocate resources to a current goal than to save resources for a future contingency.
Example 3.52 John used his last $10 for gas instead of buying an extra pair of wiper blades.
Likelihood. The more likely the successful achievement of a goal, the higher its importance. That is, an agent should devote more resources to those goals that are likely to succeed, than to less promising endeavors. Note that this dimension is related to the satisfaction conditions for a goal.
Example 3.53 John decided to study pre-law in college because he felt he had a better chance of getting into law school than into medical school.
Consequences and Outcomes. The consequences of a goal clearly affect the importance of the goal. Goals that have important consequences, either positive or negative, are themselves important.
Example 3.54 John ran a red light on his way to the hospital.
Deadline. All other things being equal, goals with impending deadlines are more important than goals with distant deadlines.
Example 3.55 John planned his Thanksgiving vacation before he started thinking about his Christmas vacation.
Frequency and Persistence. More frequent goals are more important than less frequent goals, as they require a greater allocation of resources. Goals of short duration may also be more important than more persistent goals, since they may be more pressing.
Example 3.56 John dipped into his five-year-old’s college fund to pay the mortgage.
Age. As most goals age, they grow in importance. That is, the longer that you neglect a goal, the harder it becomes to ignore. Goals such as hunger, sleep, paying taxes, and going to work all become more pressing as they go unattended.
Example 3.57 John went to the dentist once every ten years, whether he needed to or not.
Contexts and Triggers. It is hard to arrive at a general association between importance and triggering conditions for goals. We believe these to be largely independent dimensions.
Adopted Goals. As discussed in chapter 5, the importance of an adopted goal is a function of the importance of the underlying relationship. A request by your boss is generally of higher importance than a request by your secretary.
Example 3.58 John helped his mother move into her new apartment.
Serious vs. Frivolous. Serious goals are more important than frivolous goals. We view serious goals as those with greater long-term consequences. Frivolous goals provide short-term satisfaction.
Example 3.59 John wanted to go to the rock festival on Saturday, but his father had threatened to ground him if John did not pass his English exam.
Substitution..The fewer substitutes for a goal, the greater its importance. If there are many substitutes available for a goal, then the agent has the option of picking the one that requires the fewest resources. By restricting the set of options, there is an implicit increase in the resources required to achieve the goal, and an implicit increase in the goal’s importance.
Example 3.60 John did not take good care of his car since he usually walked everywhere.
Example 3.61 Mary had to commute 100 miles a day, so she was careful to maintain her car thoroughly.
The maintenance of John’s car is less important to him since he has a reliable alternative. Mary, we assume, has no choice, so she must make sure that her car works.
We need to keep in mind the underlying reason for examining and classifying the various types of goals. Our main purpose is to develop a model of how a planner chooses among disparate goals. How does a person decide what to do next?
At the heart of this process is a method for choosing or deciding between goals. The planner must be able to compare two goals. We might also describe this process in terms of plans, rather than goals. For the purposes of this discussion, we can restrict our analysis to binary decision problems — the agent must choose between two mutually exclusive outcomes.
The importance of goals provides a convenient dimension for modeling such judgments. Given a choice between pursuing two different goals, the planner will choose the one of higher importance — all other things being equal.
For example, above we posited tastes and preferences to be a type of goal. This model can then account for decisions based on such goals.
Example 3.62 John selected a red wine to go with dinner.
Example 3.63 Mary liked Cary Grant better than Robert Redford.
One might argue that these types of judgments could effectively be modeled using simple ordinal scales. For example, Cary Grant might be a 10 on Mary’s leading man scale, whereas Robert Redford might rate an 8. We could create similar scales for making other judgments.
In fact, many computer programs have made wide use of scales to account for a range of cognitive phenomena. For example, MARGIE [Schank, 1975a] used scales to represent states such as health and emotional well-being. Psychologists have used scales in experiments to provide a qualitative measure of people’s subjective judgments [Wish et al., 1976].
Despite the descriptive value and accepted utility of scales in cognitive modeling, no researcher defends the use of scales as a theoretical entity. No one claims that people have scales inside their heads. Scales are a kluge.
We submit, however, that goal importance can provide a principled way of implementing the functionality found in scales. Goal hierarchies implicitly embody a range of scales that may be used for comparisons. (We note that most adherents to scales advocated ordinal scales. Thus, if X rated a 4, and Y ranked as 2, we could conclude that X was better than Y, but not that X was twice as good as Y.)
For example, consider the scale HEALTH. In MARGIE, HEALTH(10) represented the peak of health, HEALTH(-5) suggested illness, and death was indicated by HEALTH(-10). In a goal hierarchy, we would assume that a person would have a range of goals relating to health — mostly P-goals. If all these goals were satisfied, the person would be a “10” on the HEALTH scale. If some subset of these goals were not satisfied, the person would rank lower on the scale. The actual ranking would depend on the quantity and importance of the unmet health goals. A cold might drop you 3 or 4 points, affecting relatively few low-level health goals. AIDS, on the other hand, would hit you much harder, affecting more high-level health goals.
We suggest that goal importance provides a principled mechanism for modeling decision making and goal comparison. The type of comparisons made previously with various scales emerges as a natural consequence of an importance-based hierarchy of goals. We wish to emphasize that our desire is not to provide a mechanism for inferring scales from goals. Rather, we suggest that since scales are not an intrinsic theoretical construct, we should choose to avoid them altogether by achieving the same functionality through the use of goal hierarchies — a construct that we postulate on theoretical grounds.
Given a rich set of goals, goal conflict [Wilensky, 1978] is inevitable and ubiquitous. Resources are limited. However, the principle of importance provides a reasonable method for resolving conflicts. The notion is that the agent faced with a conflict will pursue the more important goal at the expense of the less important goal.
It is of course not always simple to detect goal conflicts in the first place. However, to be more precise, we should emphasize that goal conflict is usually relative to a given plan. That is, an agent will encounter a goal conflict in the course of pursuing a plan which achieves one goal while blocking the achievement of another goal. This usually involves a perishable resource, such as time or money. (We describe other types of resources in the next chapter.) Limited resources result in goal conflict. Thus, goal conflict is not absolute or inevitable, but rather relative to specific plans and resources. We can then suggest several ways to go about resolving goal conflicts, in general. Consider a case in which an agent needs to finish writing a letter and eat lunch.
Example 3.64 John ate lunch at his desk while writing the letter to his client.
Example 3.65 John arranged a luncheon meeting with his client.
Example 3.66 John sent a short note to the client before going to lunch.
Example 3.67 John skipped lunch to finish writing his letter.
Goal conflict can be viewed as a central problem of human existence: You cannot always get what you want. By assigning priorities or importance to our goals, we have adapted to this key evolutionary challenge. If there were no goal conflict, there would be little reason to have a goal hierarchy. We would merely develop a plan to satisfy every goal. Most AI planning programs make this simplifying assumption.
Our knowledge of goals — both our own goals and those of others — comes from experience. One kind of knowledge that we have is a set of stereotypical goals for different kinds of people, as well as for people in general.
We need this kind of knowledge both for planning and understanding. Here are some examples of the type of default goal knowledge that people have.
Infant: Learn to speak, satisfy hunger, thirst.
Child: Play games, learn to read, make friends.
Teenager: Play sports, date, learn to take SAT’s.
College student: Choose a major, choose a career.
Adult: Get a job, get married, save money, find apartment.
Married adult: Save more money, get a house, have children.
Parent: Save more money, send children to school.
Grandparent: Retire, visit grandchildren.
The point is simple. Just as we have scriptal knowledge about what goes on in restaurants and dentist offices [Schank and Abelson, 1977], we have default knowledge about stereotypical sets of goals and priorities for different types of people. The preceding list followed a chronological dimension. It would be easy to create similar lists for other categories, such as vocation or culture. For example, there would be sets of normative goals associated with doctors, plumbers, and bankers, and with Americans, Japanese, and Indians. This type of knowledge, like scripts, is a product of our experience. The cases in our experience can provide us with these generalizations.
Schank and Lebowitz (1979) discussed normative inferences based on a variety of personal characteristics, including profession, familial relationship, life style, habitual activities, intellectual affiliation, social affiliation, and attitudes. These same dimensions will imply normative goals.
This knowledge of stereotypical goals provides us with expectations about behavior of other people. We also have a similar kind of knowledge about certain inanimate objects — functional objects. We can ascribe to these objects properties similar to goals, but in this case certainly not the goals of the objects themselves. Rather, we can infer the goal of the designer or user of the object.
We offer two examples of designed objects: a car and a computer. For each, there are a number of default goals.
Car: Move when the accelerator is depressed, stop when the brake is engaged, recharge the battery when the engine is running, channel the exhaust through the tailpipe and not the passenger compartment, move the wheels in the direction indicated by the steering wheel, alert driver of speed, engine temperature, and fluid levels.
Computer: Monitor input on all devices, display output on appropriate devices, check authorization of users making file read/write requests, verify passwords of users, monitor disk usage, execute user programs.
We are not suggesting that automobiles or computers themselves have goals. Rather, we mean to state that these functional objects embody the goals of their designers. Their processes reflect a purpose.
Knowledge of such embedded goals supports the two complementary tasks of design and understanding. In designing an object, such as a car, an agent must be cognizant of the intentions of the user. The agent must then form the object to fit this function. For example, it should be relatively easy to add gas to a car, since that task is a fairly frequent one, but changing the oil does not need to be as convenient, since that task is relatively infrequent.
On the other hand, the user of a functional object can understand its behavior better by reflecting on the designer’s intentions. The owner of a car can infer that the shiny dashboard controls are made of chrome or plastic, rather than silver or platinum, since those latter materials would significantly increase the manufacturing cost.
In summary, we have proposed a wide variety of goals that people may choose to pursue. We suggest that most people will have more goals than resources, and thus will be forced to make choices among their goals. In making decisions, agents will rely on a vast, detailed knowledge of goals — both their own and the goals that one might expect others to have. The study of goals should provide a principled way to model a range of human cognitive behavior. In the next chapter we discuss resources in detail, and in the following chapter, we present our model of goal-based interpersonal relationships.